Merge pull request #44235 from xxhdx1985126/wip-onode-omap-hint-optimization

crimson/os/seastore: avoid onode/omap laddr hint conflicts as much as possible

Reviewed-by: Samuel Just <sjust@redhat.com>
Reviewed-by: Yingxin Cheng <yingxin.cheng@intel.com>
Reviewed-by: Chunmei Liu <chunmei.liu@intel.com>

src/common/options/crimson.yaml.in

@@ -65,3 +65,13 @@ options:
   level: dev
   desc: The record fullness threshold to flush a journal batch
   default: 0.95
+- name: seastore_default_max_object_size
+  type: uint
+  level: dev
+  desc: default logical address space reservation for seastore objects' data
+  default: 16777216
+- name: seastore_default_object_metadata_reservation
+  type: uint
+  level: dev
+  desc: default logical address space reservation for seastore objects' metadata
+  default: 16777216

src/crimson/common/interruptible_future.h

@@ -10,8 +10,11 @@
 
 #include "crimson/common/log.h"
 #include "crimson/common/errorator.h"
-
+#ifndef NDEBUG
 #define INTR_FUT_DEBUG(FMT_MSG, ...) crimson::get_logger(ceph_subsys_).trace(FMT_MSG, ##__VA_ARGS__)
+#else
+#define INTR_FUT_DEBUG(FMT_MSG, ...)
+#endif
 
 // The interrupt condition generally works this way:
 //

src/crimson/os/seastore/object_data_handler.cc

@@ -15,16 +15,6 @@ namespace {
 }
 
 namespace crimson::os::seastore {
-
-/**
- * MAX_OBJECT_SIZE
- *
- * For now, we allocate a fixed region of laddr space of size MAX_OBJECT_SIZE
- * for any object.  In the future, once we have the ability to remap logical
- * mappings (necessary for clone), we'll add the ability to grow and shrink
- * these regions and remove this assumption.
- */
-static constexpr extent_len_t MAX_OBJECT_SIZE = 16<<20;
 #define assert_aligned(x) ceph_assert(((x)%ctx.tm.get_block_size()) == 0)
 
 using context_t = ObjectDataHandler::context_t;
@@ -261,9 +251,9 @@ ObjectDataHandler::write_ret ObjectDataHandler::prepare_data_reservation(
   extent_len_t size)
 {
   LOG_PREFIX(ObjectDataHandler::prepare_data_reservation);
-  ceph_assert(size <= MAX_OBJECT_SIZE);
+  ceph_assert(size <= max_object_size);
   if (!object_data.is_null()) {
-    ceph_assert(object_data.get_reserved_data_len() == MAX_OBJECT_SIZE);
+    ceph_assert(object_data.get_reserved_data_len() == max_object_size);
     DEBUGT("reservation present: {}~{}",
            ctx.t,
            object_data.get_reserved_data_base(),
@@ -272,14 +262,14 @@ ObjectDataHandler::write_ret ObjectDataHandler::prepare_data_reservation(
   } else {
     DEBUGT("reserving: {}~{}",
            ctx.t,
-           ctx.onode.get_hint(),
-           MAX_OBJECT_SIZE);
+           ctx.onode.get_data_hint(),
+           max_object_size);
     return ctx.tm.reserve_region(
       ctx.t,
-      ctx.onode.get_hint(),
-      MAX_OBJECT_SIZE
-    ).si_then([&object_data](auto pin) {
-      ceph_assert(pin->get_length() == MAX_OBJECT_SIZE);
+      ctx.onode.get_data_hint(),
+      max_object_size
+    ).si_then([max_object_size=max_object_size, &object_data](auto pin) {
+      ceph_assert(pin->get_length() == max_object_size);
       object_data.update_reserved(
 	pin->get_laddr(),
 	pin->get_length());

src/crimson/os/seastore/object_data_handler.h

@@ -50,6 +50,8 @@ class ObjectDataHandler {
 public:
   using base_iertr = TransactionManager::base_iertr;
 
+  ObjectDataHandler(uint32_t mos) : max_object_size(mos) {}
+
   struct context_t {
     TransactionManager &tm;
     Transaction &t;
@@ -104,6 +106,16 @@ private:
     context_t ctx,
     object_data_t &object_data,
     extent_len_t size);
+private:
+  /**
+   * max_object_size
+   *
+   * For now, we allocate a fixed region of laddr space of size max_object_size
+   * for any object.  In the future, once we have the ability to remap logical
+   * mappings (necessary for clone), we'll add the ability to grow and shrink
+   * these regions and remove this assumption.
+   */
+  const uint32_t max_object_size = 0;
 };
 
 }

src/crimson/os/seastore/onode.h

@@ -52,12 +52,29 @@ class Onode : public boost::intrusive_ref_counter<
   Onode,
   boost::thread_unsafe_counter>
 {
+protected:
+  virtual laddr_t get_hint() const = 0;
+  const uint32_t default_metadata_offset = 0;
+  const uint32_t default_metadata_range = 0;
 public:
+  Onode(uint32_t ddr, uint32_t dmr)
+    : default_metadata_offset(ddr),
+      default_metadata_range(dmr)
+  {}
 
   virtual const onode_layout_t &get_layout() const = 0;
   virtual onode_layout_t &get_mutable_layout(Transaction &t) = 0;
   virtual ~Onode() = default;
-  virtual laddr_t get_hint() const = 0;
+
+  laddr_t get_metadata_hint() const {
+    assert(default_metadata_offset);
+    assert(default_metadata_range);
+    return get_hint() + default_metadata_offset +
+      ((uint32_t)std::rand() % default_metadata_range);
+  }
+  laddr_t get_data_hint() const {
+    return get_hint();
+  }
 };
 
 

src/crimson/os/seastore/onode_manager/staged-fltree/fltree_onode_manager.cc

@@ -27,7 +27,10 @@ FLTreeOnodeManager::get_onode_ret FLTreeOnodeManager::get_onode(
       DEBUGT("no entry for {}", trans, hoid);
       return crimson::ct_error::enoent::make();
     }
-    auto val = OnodeRef(new FLTreeOnode(cursor.value()));
+    auto val = OnodeRef(new FLTreeOnode(
+	default_data_reservation,
+	default_metadata_range,
+	cursor.value()));
     return get_onode_iertr::make_ready_future<OnodeRef>(
       val
     );
@@ -43,10 +46,13 @@ FLTreeOnodeManager::get_or_create_onode(
   return tree.insert(
     trans, hoid,
     OnodeTree::tree_value_config_t{sizeof(onode_layout_t)}
-  ).si_then([&trans, &hoid, FNAME](auto p)
+  ).si_then([this, &trans, &hoid, FNAME](auto p)
               -> get_or_create_onode_ret {
     auto [cursor, created] = std::move(p);
-    auto val = OnodeRef(new FLTreeOnode(cursor.value()));
+    auto val = OnodeRef(new FLTreeOnode(
+	default_data_reservation,
+	default_metadata_range,
+	cursor.value()));
     if (created) {
       DEBUGT("created onode for entry for {}", trans, hoid);
       val->get_mutable_layout(trans) = onode_layout_t{};

src/crimson/os/seastore/onode_manager/staged-fltree/fltree_onode_manager.h

@@ -37,7 +37,14 @@ struct FLTreeOnode final : Onode, Value {
   FLTreeOnode& operator=(const FLTreeOnode&) = delete;
 
   template <typename... T>
-  FLTreeOnode(T&&... args) : Value(std::forward<T>(args)...) {}
+  FLTreeOnode(uint32_t ddr, uint32_t dmr, T&&... args)
+    : Onode(ddr, dmr),
+      Value(std::forward<T>(args)...) {}
+
+  template <typename... T>
+  FLTreeOnode(T&&... args)
+    : Onode(0, 0),
+      Value(std::forward<T>(args)...) {}
 
   struct Recorder : public ValueDeltaRecorder {
     Recorder(bufferlist &bl) : ValueDeltaRecorder(bl) {}
@@ -102,12 +109,23 @@ struct FLTreeOnode final : Onode, Value {
 
 using OnodeTree = Btree<FLTreeOnode>;
 
+using crimson::common::get_conf;
+
 class FLTreeOnodeManager : public crimson::os::seastore::OnodeManager {
   OnodeTree tree;
 
+  uint32_t default_data_reservation = 0;
+  uint32_t default_metadata_offset = 0;
+  uint32_t default_metadata_range = 0;
 public:
   FLTreeOnodeManager(TransactionManager &tm) :
-    tree(NodeExtentManager::create_seastore(tm)) {}
+    tree(NodeExtentManager::create_seastore(tm)),
+    default_data_reservation(
+      get_conf<uint64_t>("seastore_default_max_object_size")),
+    default_metadata_offset(default_data_reservation),
+    default_metadata_range(
+      get_conf<uint64_t>("seastore_default_object_metadata_reservation"))
+  {}
 
   mkfs_ret mkfs(Transaction &t) {
     return tree.mkfs(t);

src/crimson/os/seastore/seastore.cc

@@ -68,6 +68,8 @@ public:
   }
 };
 
+using crimson::common::get_conf;
+
 SeaStore::SeaStore(
   const std::string& root,
   MDStoreRef mdstore,
@@ -80,7 +82,9 @@ SeaStore::SeaStore(
     segment_manager(std::move(sm)),
     transaction_manager(std::move(tm)),
     collection_manager(std::move(cm)),
-    onode_manager(std::move(om))
+    onode_manager(std::move(om)),
+    max_object_size(
+      get_conf<uint64_t>("seastore_default_max_object_size"))
 {
   register_metrics();
 }
@@ -459,7 +463,7 @@ SeaStore::read_errorator::future<ceph::bufferlist> SeaStore::read(
 	size - offset :
 	std::min(size - offset, len);
 
-      return ObjectDataHandler().read(
+      return ObjectDataHandler(max_object_size).read(
         ObjectDataHandler::context_t{
           *transaction_manager,
           t,
@@ -509,7 +513,7 @@ SeaStore::get_attr_errorator::future<ceph::bufferlist> SeaStore::get_attr(
       }
       return _omap_get_value(
         t,
-        layout.xattr_root.get(onode.get_hint()),
+        layout.xattr_root.get(onode.get_metadata_hint()),
         name);
     }
   ).handle_error(crimson::ct_error::input_output_error::handle([FNAME] {
@@ -607,7 +611,8 @@ SeaStore::omap_get_values(
     "omap_get_values",
     op_type_t::OMAP_GET_VALUES,
     [this, keys](auto &t, auto &onode) {
-      omap_root_t omap_root = onode.get_layout().omap_root.get(onode.get_hint());
+      omap_root_t omap_root = onode.get_layout().omap_root.get(
+	onode.get_metadata_hint());
       return _omap_get_values(
 	t,
 	std::move(omap_root),
@@ -685,7 +690,7 @@ SeaStore::_omap_list_ret SeaStore::_omap_list(
   const std::optional<std::string>& start,
   OMapManager::omap_list_config_t config) const
 {
-  auto root = omap_root.get(onode.get_hint());
+  auto root = omap_root.get(onode.get_metadata_hint());
   if (root.is_null()) {
     return seastar::make_ready_future<_omap_list_bare_ret>(
       true, omap_values_t{}
@@ -1068,7 +1073,7 @@ SeaStore::tm_ret SeaStore::_write(
   return seastar::do_with(
     std::move(_bl),
     [=, &ctx, &onode](auto &bl) {
-      return ObjectDataHandler().write(
+      return ObjectDataHandler(max_object_size).write(
         ObjectDataHandler::context_t{
           *transaction_manager,
           *ctx.transaction,
@@ -1089,13 +1094,13 @@ SeaStore::_omap_set_kvs(
 {
   return seastar::do_with(
     BtreeOMapManager(*transaction_manager),
-    omap_root.get(onode->get_hint()),
+    omap_root.get(onode->get_metadata_hint()),
     [&, keys=std::move(kvs)](auto &omap_manager, auto &root) {
       tm_iertr::future<> maybe_create_root =
         !root.is_null() ?
         tm_iertr::now() :
         omap_manager.initialize_omap(
-          t, onode->get_hint()
+          t, onode->get_metadata_hint()
         ).si_then([&root](auto new_root) {
           root = new_root;
         });
@@ -1146,13 +1151,13 @@ SeaStore::tm_ret SeaStore::_omap_rmkeys(
 {
   LOG_PREFIX(SeaStore::_omap_rmkeys);
   DEBUGT("{} {} keys", *ctx.transaction, *onode, keys.size());
-  auto omap_root = onode->get_layout().omap_root.get(onode->get_hint());
+  auto omap_root = onode->get_layout().omap_root.get(onode->get_metadata_hint());
   if (omap_root.is_null()) {
     return seastar::now();
   } else {
     return seastar::do_with(
       BtreeOMapManager(*transaction_manager),
-      onode->get_layout().omap_root.get(onode->get_hint()),
+      onode->get_layout().omap_root.get(onode->get_metadata_hint()),
       std::move(keys),
       [&ctx, &onode](
 	auto &omap_manager,
@@ -1198,7 +1203,7 @@ SeaStore::tm_ret SeaStore::_truncate(
   LOG_PREFIX(SeaStore::_truncate);
   DEBUGT("onode={} size={}", *ctx.transaction, *onode, size);
   onode->get_mutable_layout(*ctx.transaction).size = size;
-  return ObjectDataHandler().truncate(
+  return ObjectDataHandler(max_object_size).truncate(
     ObjectDataHandler::context_t{
       *transaction_manager,
       *ctx.transaction,

src/crimson/os/seastore/seastore.h

@@ -305,6 +305,7 @@ private:
   TransactionManagerRef transaction_manager;
   CollectionManagerRef collection_manager;
   OnodeManagerRef onode_manager;
+  const uint32_t max_object_size = 0;
 
   using tm_iertr = TransactionManager::base_iertr;
   using tm_ret = tm_iertr::future<>;

src/test/crimson/seastore/onode_tree/test_fltree_onode_manager.cc

@@ -31,15 +31,15 @@ struct onode_item_t {
   void initialize(Transaction& t, Onode& value) const {
     auto& layout = value.get_mutable_layout(t);
     layout.size = size;
-    layout.omap_root.update(omap_root_t(id, cnt_modify, value.get_hint()));
+    layout.omap_root.update(omap_root_t(id, cnt_modify, value.get_metadata_hint()));
     validate(value);
   }
 
   void validate(Onode& value) const {
     auto& layout = value.get_layout();
     ceph_assert(laddr_t(layout.size) == laddr_t{size});
-    ceph_assert(layout.omap_root.get(value.get_hint()).addr == id);
-    ceph_assert(layout.omap_root.get(value.get_hint()).depth == cnt_modify);
+    ceph_assert(layout.omap_root.get(value.get_metadata_hint()).addr == id);
+    ceph_assert(layout.omap_root.get(value.get_metadata_hint()).depth == cnt_modify);
   }
 
   void modify(Transaction& t, Onode& value) {

src/test/crimson/seastore/test_object_data_handler.cc

@@ -11,6 +11,10 @@ using namespace crimson;
 using namespace crimson::os;
 using namespace crimson::os::seastore;
 
+#define MAX_OBJECT_SIZE (16<<20)
+#define DEFAULT_OBJECT_DATA_RESERVATION (16<<20)
+#define DEFAULT_OBJECT_METADATA_RESERVATION (16<<20)
+
 namespace {
   [[maybe_unused]] seastar::logger& logger() {
     return crimson::get_logger(ceph_subsys_test);
@@ -22,6 +26,7 @@ class TestOnode final : public Onode {
   bool dirty = false;
 
 public:
+  TestOnode(uint32_t ddr, uint32_t dmr) : Onode(ddr, dmr) {}
   const onode_layout_t &get_layout() const final {
     return layout;
   }
@@ -58,7 +63,7 @@ struct object_data_handler_test_t:
 	offset,
 	len));
     with_trans_intr(t, [&](auto &t) {
-      return ObjectDataHandler().write(
+      return ObjectDataHandler(MAX_OBJECT_SIZE).write(
         ObjectDataHandler::context_t{
           *tm,
           t,
@@ -81,7 +86,7 @@ struct object_data_handler_test_t:
 	0,
 	size - offset);
       with_trans_intr(t, [&](auto &t) {
-        return ObjectDataHandler().truncate(
+        return ObjectDataHandler(MAX_OBJECT_SIZE).truncate(
           ObjectDataHandler::context_t{
             *tm,
             t,
@@ -100,7 +105,7 @@ struct object_data_handler_test_t:
 
   void read(Transaction &t, objaddr_t offset, extent_len_t len) {
     bufferlist bl = with_trans_intr(t, [&](auto &t) {
-      return ObjectDataHandler().read(
+      return ObjectDataHandler(MAX_OBJECT_SIZE).read(
         ObjectDataHandler::context_t{
           *tm,
           t,
@@ -132,7 +137,9 @@ struct object_data_handler_test_t:
   }
 
   seastar::future<> set_up_fut() final {
-    onode = new TestOnode{};
+    onode = new TestOnode(
+      DEFAULT_OBJECT_DATA_RESERVATION,
+      DEFAULT_OBJECT_METADATA_RESERVATION);
     known_contents = buffer::create(4<<20 /* 4MB */);
     memset(known_contents.c_str(), 0, known_contents.length());
     size = 0;